ESTIMATING THE IDEOLOGY OF PRIMARY ELECTORATES
Rachel A. Surminsky
A thesis submitted to the faculty of the University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Master of Arts in the Department of Political Science, concentration in American Politics.
Chapel Hill 2018
© 2018
ABSTRACT
RACHEL A. SURMINSKY: Estimating the Ideology of Primary Electorates (Under the direction of Sarah A. Treul.)
TABLE OF CONTENTS
LIST OF TABLES . . . v
LIST OF FIGURES . . . vi
INTRODUCTION . . . 1
LIST OF FIGURES
Figure
1 Republican Party Primary Electorate Ideology by Congressional District . . 16 2 Democrat Party Primary Electorate Ideology by Congressional District . . 16 3 Democratic Party Primary Electorate vs. General Electorate Ideological
Estimates . . . 18 4 Republican Party Primary Electorate vs. General Electorate Ideological
INTRODUCTION
More so today than in the past, congressional districts favor one party over the other (Car-son, Crespin, Eaves and Wanless 2012). With increased frequency, voters are casting parti-san ballots (Jacobson 2009). In the 2016 general election, just thirteen House seats switched party control and 96% of general election voters selected a presidential and congressional candidate from the same party (Carson, Roberts and Surminsky 2017). General elections have become increasingly consistent, producing predictably partisan outcomes. Primary elections, however, have become less predictable. Since 2010, the number of unopposed primary elections has dropped dramatically, demonstrating a shift towards greater intra-party competition (Porter and Treul 2018). Incumbents from safely partisan districts who would presumably win in the general election now fear losing to an in-party challenger in the primary. For some members of Congress it seems the threat of losing reelection has shifted from the general election to the primary election.
Competitive primaries can put an incumbent in a tough position because she must ap-peal to two different voting electorates. She must please voters in the primary election to secure her party’s nomination before running in the general election where she must please a different set of voters. In contemporary elections, the average primary voter is thought to be more ideologically extreme1 than the average general election voter. Primary constituents are more likely to reward or punish an incumbent for her voting record than are general election constituents (Sides, Tausanovitch, Vavreck and Warshaw Forthcoming). Ideologi-cally extreme voters encourage more ideological challengers to enter at the primary stage, elevating the likelihood of intra-party competition (Butler 2009; Maestas, Fulton, Maisel and Stone 2006). Ideological challengers may better fit the preferences of the primary
torate, swaying voters away from the incumbent. Faced with divergent constituencies, an incumbent may choose to appeal more heavily to her ideologically extreme primary elec-torate instead of the moderate general elecelec-torate. To the extent that many members of Congress are elected from safe districts, an uncertain or vulnerable primary election may make pleasing primary voters a higher priority. In turn, if an incumbent is trying to win over ideologically extreme primary voters, she will engage in more ideologically extreme behavior.
This theory for an ideologically extreme primary electorate pulling incumbents away from the median voter is intuitive. Current research on ideological candidate emergence and polarized member behavior fit with this conception of primary elections (Thomsen 2014; Hall 2015; Jewitt and Treul 2018). Anecdotally, examples like the primary election defeat of Richard Lugar (R-IN) fall in line with this argument as well. However, our ability to directly investigate the influence of ideologically extreme primary voters is limited by the lack of better measures for each district’s primary election constituency (Brady, Han and Pope 2007).
Current measures of primary constituency ideology may not accurately capture district-by-district heterogeneity in the ideological distribution of constituents. The limitations of these measures may account for mixed findings on the influence, or even existence, of ideologically extreme primary election voters (for a review see Hirano, Jr., Ansolabehere and Hansen 2010; Brady, Han and Pope 2007; Boatright 2014; Clinton 2006; Sides et al. Forthcoming; Abramowitz 2008). If we want to be able to definitively demonstrate that primary election voters are more ideological and test claims about their impact on candidate behavior, we need an new, more direct measure for constituency ideology.
general electorate ideology within and across congressional districts.
Examining primary and general electorates for U.S. House districts in the 2012 election, I show that primary voters are more ideologically extreme than general election voters. I also find evidence that there is variation in the extent to which primary and general election voters ideologically diverge. These findings support my theory that the ideological distri-bution of primary voters is heterogeneous across districts. While primary voters may be more extreme on average, my estimates suggest that some districts have electorates that are very ideologically similar. Comparing between parties, I additionally demonstrate that Republican primary voters are often more ideologically extreme than Democratic primary voters.
My measure demonstrates the necessity for a district-level examination of electorate ideology. Candidate behavior is conditional on district context (Fenno 1978; Maestas et al. 2006). Our theories about the influence of primary voters on candidate behavior must be conditional as well, accounting for ideological heterogeneity across districts. Future work should tailor expectations for candidate behavior to the specific ideological distribution of voters in each candidate’s district.
A Need for Electorate-Level Estimates of Ideology
Several studies use survey data to directly measure the ideological predispositions of pri-mary voters. Clinton (2006) and Butler (2009) use data from the Congressional Cooperative Elections Survey (CCES) to show that same-party constituency preferences motivate candi-dates to move their positions towards the extreme. Jacobson (2009) looks at the ideological predispositions of primary voters using data from the American National Election Survey (ANES), finding them more extreme than general election voters. These findings point to the presence of ideologically extreme primaryvotersbut do not make conclusions about in which districts we should expect to find ideologically extreme primaryelectorates.2
A sizable literature argues that incumbent behavior is conditional on her constituency. Fenno’s (1978) seminal work finds that the representative-constituent relationship is district-specific. Further work suggest an incumbent will use updated information about her con-stituency to shift her behavior in the ideological direction of constituents (Rabinowitz and Macdonald 1989; Kousser, Lewis and Masket 2007; Fleisher and Bond 2004; Sulkin 2005; Jewitt and Treul 2018; Clinton 2006). Political economy models of two-stage electoral competition in primary elections predict that an incumbent will diverge from her party con-ditional on the ideological clustering of her primary constituency (Aranson and Ordeshook 1972; Owen and Grofman 2006; Grofman 2004).3 In theory, the electoral clout of ideo-logical primary voters should increase with the quantity of ideoideo-logically extreme primary voters in a given district. To understand the electoral influence of primary voters, it is not enough to know that primary voters can be extreme. We need to know in which dis-tricts these voters tend to be more extreme, forming expectations about incumbent behavior accordingly.
A similar argument can be made for candidate emergence. Strategic, politically expe-rienced candidates are more likely to run when national and local conditions are favorable,
2Several analyses have looked at the ideological extremity of primary voters at the national or state level, finding that on aggregate primary voters are not that much more extreme than general election voters (Sides et al. Forthcoming). These explorations provide interesting insight but mask variance in the ideological dis-tribution of voters across districts, which could solicit a polarizing effect on incumbent behavior
acutely aware of the costs and benefits to running (Jacobson 1989; Hetherington, Larson and Globetti 2003; Maestas et al. 2006). Theories of candidate emergence in the general election have been applied at the primary level, finding similar results: candidates run in primary elections when district-level conditions are the most favorable (Thomsen 2014; Porter and Treul 2018). Without a way to distinguish one primary electorate from another, we cannot pinpoint those districts where we would expect an extreme primary electorate to motivate ideological candidate emergence. If candidate behavior is conditional on a particular district’s ideological composition and the ideological extremity of the primary electorate varies across districts, a direct estimate for electorate ideology is necessary to test for a primary constituency effect on candidate emergence.
Current Conceptions of Primary Voter Ideology
Conventional methods for estimating public opinion use data on individual-level voter pref-erences from national surveys. Survey respondents can be disaggregated into smaller sub-samples, for instance by county or congressional district, to estimate voter ideology at the subnational level (Miller and Stokes 1963; Gelman and Little 1997; Leemann and Wasser-fallen 2017). However, there is very little survey data on primary elections. National sur-veys such as the American National Election Survey (ANES) and Cooperative Congres-sional Elections Survey (CCES) infrequently ask questions regarding primary election vote choice or participation. Additionally, sampling for these large, national-level surveys is not representative of each primary constituency for each party at the congressional dis-trict level. Once survey respondents are disaggregated into subnational units, for example the primary constituency for the Democratic party in a given district, there is a significant small-N problem. Limited survey data restricts our ability to use traditional approaches to estimate primary electorate ideology.
primary electorate. For example, to determine if primary voters influence candidate be-havior, McGhee, Masket, Shor, Rogers and McCarty (2014) compare districts with open primary institutions to more closed systems. They assume that states with exclusionary closed primaries will have more partisan, ideological primary electorates; representatives from these states should be more polarized. The authors find no evidence that incumbent behavior is more polarized in districts with closed institutions. Hill (2015) tests McGhee et al.’s (2014) assumption finding that the distribution of voter ideology within primary and general electorates does not correlate with a state’s type of primary institution. His work demonstrates that proxy measures for district ideology can mischaracterize the extremity of a district’s primary electorate.
In a similar vein, Boatright (2014) and Lawless and Pearson (2008) find little differ-ence in member behavior when comparing across different levels of primary competition, assuming that highly competitive elections should produce an incumbent behavior shift to accommodate primary voter preferences. Jewitt and Treul (2018) argue that close elec-tions may not lead to changes in incumbent behavior, instead pointing to divisive races — races that are ideological in nature — as challenges with behavior-altering consequences. Hirano et al. (2010) find no evidence of incumbent behavior change before and after the introduction of primary elections in the late 1950’s to early 1960’s. Hill and Tausanovitch (2017) demonstrate that primary electorates today are more ideological than they were in the past. No change in incumbent behavior should have been expected after the introduction of primary elections because primary constituencies in the 1960’s were not yet ideologically extreme. These types of investigations into the influence of ideologically extreme primary voters make a strong assumption, inferring that there is ideological consistency in primary electorates across most districts.4
Restricted by data availability, our capacity to investigate the electoral impacts of pri-mary voters has previously been limited to these types of approaches. Building on the groundwork laid by Leemann and Wasserfallen (2017), I employ a new method to develop
a measure for primary and general electorate ideology in both parties at the congressional district-level. I demonstrate that the ideological composition of primary and general elec-tion voters does, indeed, vary across districts. My estimates serve to further our under-standing of subnational public opinion in the United States and allow for more thorough investigations of primary constituency influence on candidate behavior.
Subnational Public Opinion Estimation: A Methodological Overview
Significant strides in the study of subnational public opinion have been made recently us-ing multi-level regression with postratification (MrP). The utility of MrP comes from its ability to produce more precise estimates of subnational public opinion than traditional dis-aggregation (Lax and Phillips 2009a). MrP up-weights specific geographies that may be undersampled in survey data and down-weights oversampled subpopulations to make esti-mates more representative. This approach developed by Gelman and Little (1997) has been applied in a variety of contexts to measure public opinion in the United States at the state and local level (Park, Gelman and Bafumi 2004; Lax and Phillips 2009b; Tausanovitch and Warshaw 2013; Warshaw and Rodden 2012).
within a state produces an estimate of public opinion at the state level.
To weight — or poststratify — predictions, fine-grain data from the U.S. census is often employed. U.S. census data is excellent for postratification because: (1) it serves as a reliable picture of the population, (2) information is broken down by various subnational units of interest for instance county or congressional district, and (3) data is provided in the form of joint distributions. This means that the census not only provides a count of the number of men in a given state but also the number of men who are 20-25 and have a college education in a given state.
The ability of MrP to precisely measure public opinion at very localized levels makes it a promising solution for estimating electorate ideology. However, joint distributions are necessary for MrP; without them public opinion predictions cannot be weighted to reflect the distribution of voter types in a given population. This stringent data requirement has made it impossible to use MrP to estimate constituency ideology. The U.S. census does not release any information on individual-level electoral participation or partisan affiliation. In other words, there are no available joint distributions for the number of men in a given con-gressional district who are 20-25, have a college education,and belong to the Republican party primary electorate.5
A recent development by Leemann and Wasserfallen (2017) relaxes the necessity of joint distributions in estimating subnational public opinion.6 This variation of MrP – known as multilevel regression with synthetic poststratification or MrsP — allows for marginal distributions to be used to impute unknown joint distributions. For example, with a known joint distribution of college-educated men in a specific district and a known
5This limitation is not exclusive to examinations of voter participation or partisanship, Warshaw and Rod-den (2012) could not use age as a predictor in the MrP model for district level public opinion on individual issue areas.
marginal distribution of registered Republicans in a specific district, MrsP can be used to impute how many college-educated men in that district are registered Republicans. This imputed joint distribution is what Leemann and Wasserfallen (2017) call a “synthetic joint distribution”. Correlations amongst empirical joint distributions of individual-level vari-ables can be used to help inform the creation of synthetic joint distributions. For example, if college educated men tend to be more Republican, this information is used to help guide our estimations. These “adjusted synthetic joint distributions” take into account dependen-cies across covariates.
Data, Measurement, & Analysis of District Ideology
I use MrsP to create a direct measure for primary and general electorate ideology. MrsP performs just as well, and in some instances better than, traditional MrP. It does not require census-level joint distributions that include a person’s electoral participation, which has prevented its application for this purpose in the past (Leemann and Wasserfallen 2017). MrsP overcomes the methodological hurdle in producing of a district-level measure for constituency ideology, but data availability is still an obstacle.
To nest survey respondents within the primary or general electorate for each district re-quires knowing if they voted in the primary election, the general election, or both. As previ-ously noted, questions about primary voter participation are highly infrequent in national-level surveys.7 Further, self-reported participation measures often over-report election turnout (Butler 2009; Sides et al. Forthcoming; Vavreck 2007). Using validated voter turnout in the CCES resolves this problem, but drastically reduces already small sub-samples of respondents nested within each party’s primary and general electorate for each district.8 The average number of CCES respondents who voted in the primary election
7For example, the American National Elections Survey only asked about primary election turnout in 1958, 1964, 1966, and 1978. The ANES asked about presidential primary participation in 1992, not congressional primary participation.
for a given party in each congressional district is approximately 15; some district primary constituencies have no respondents, a few have as many as 100.9
Instead of national surveys, I use Catalist LLC’s Validated Voter Database as a data source for MrsP estimation. Catalist aggregates voter files for all 50 states, also drawing on external data sources to build profiles of individuals. The academic subscription pro-vides a 1% randomly drawn sample of their database. My sample drawn in 2014 for the 2012 election includes 3.1 million cases, approximately 890,000 primary voters. The Catal-ist Voter Database includes demographic information, validated election participation, and party registration when available. This data from Catalist, LLC acts as a good substitute to survey data, including similar information with a much larger sample size of primary election voters. However, using the data involves making some assumptions, discussed in the following section.
Estimating Voter Ideology
To build estimates of constituency ideology by party for each congressional district, I employ a hierarchical linear model. It is similar to those adopted in previous studies exam-ining voter ideological extremity (Hill and Tausanovitch 2017; Sides et al. Forthcoming). I regress an individual’s ideological extremity on a standard set of demographic characteris-tics including age, gender, education level, and race. I also include indicators variables for state and congressional district along with dummy variables for 2012 Democratic primary participation and 2012 Republican primary participation.
Ideologyi =β0+αrace[i] +αgender[i] +αedu[i]+βdemprimary[i]+βrepprimary[i]
+αage[i]+αstate[i]+αdistrict[i]+ε[i]
α·∼ N(0, σ·2)
αdistrict∼ N(γ1PresVote, σ2district)
ε[i]∼ N(0, σy)
All predictors are modeled using random effects except party primary participation modeled using fixed effects.10 I let my model intercept vary by congressional district and state. Random effects are drawn from a zero mean normal distribution, though the dis-trict level covariate is drawn from a distribution centered on the the Democratic presidential vote share for that district.11 A discussion of the measurement for each variable in the model is as follows.
Voter Ideological Extremity
To help clients analyze the electorate and target constituents, Catalist, LLC builds pre-dictive scores using their Voter Validated Data for a respondent’s ideological extremity. This synthetic score is scaled from 0-100 with 0 being the most conservative and 100 being the most liberal. The variable is constructed using more than 150 covariates and is gener-ally accepted as a reliable measure for relative ideological extremity between individuals (Hersh 2015).
10I use fixed effects because I assume Democratic and Republic primary voter ideological extremity is not drawn from a common distribution. There is no borrowing of information across groups to inform an individual’s level of ideological extremity. The ideology of Democratic and Republican primary voters will be fundamentally different.
Age
The age variable in the Catalist database is respondent’s age at the time of the next general election; subtracting two from this number gives a respondent’s age at the time of the 2012 election. Only individuals in the voting age population are included in the analysis. Data on respondent age provided by Catalist is drawn from state voter files.
Gender
This dummy variable equals 1 if the respondent is female, 0 otherwise. Data on respon-dent gender provided by Catalist is also drawn from state voter files.
Education
State voter files do not include information about an individual’s education level. Based on geographic information, consumer information, and other covariates, Catalist, LLC cre-ates a propensity score for a respondent’s likelihood to have a certain level of educational attainment. Catalist includes six potential bins for respondent education, I collapse this into three categories: Not High School Graduate, High School Graduate, and Bachelor’s De-gree/Postgraduate.12 Comparing across these three categories, the respondent is coded for having the educational attainment of whichever propensity score is the highest.
Race
The race variable provided by Catalist includes race and ethnicity categories different from those included in U.S. census data. In order to weight model predictions in the pos-tratification stage of MrsP, these categories are binned to match those in the census. Race categories include Caucasian, Black, Hispanic, and Other. Data is drawn from state voter
files and commercial sources.13
MrsP Poststratification
I use this model to calculate predicted values of ideological extremity for all voter types in the Catalist Validated Voter Database. To make these predictions representative at the district-level, they are weighted with the percentage of that voter type within the actual population of each congressional district. The U.S. census provides the joint distributions for all but two individual-level predictors in the model. The census factfinder does not in-clude age or voter participation in its joint distributions. For these demographic character-istics I impute joint distributions using marginal distributions. To account for correlations amongst individual-level predictors, I use adjusted synthetic joint distributions instead of simple synthetic joint distributions.
To produce an adjusted synthetic joint distributions for age, I use the marginal distribu-tion for age provided by the U.S. census. However, the U.S. census provides no informadistribu-tion about voter participation in the primary or general election. In lieu of census data, I charac-terize the marginal distribution for the Republican (Democratic) primary electorate as the total number of voters who participated in the Republican (Democratic) primary. For the general electorate, I use the total number of voters who voted for the Republican candidate and the total number of voters who voted for the Democratic candidate as my marginal distribution.
Using voter turnout as my marginal distribution could be problematic for several rea-sons. First, if a race is unopposed, there is no recorded vote total in that party’s primary. Therefore, no marginal distribution exists for voter turnout and no ideological estimate can be produced in that district for the party’s primary constituency. On one hand, this could indicate that a representative matches her constituency well; on the other, it may simply be that no challenger decided to run. Regardless, this limits the explanatory power of my estimates.
Second, voter turnout in elections fluctuates year-to-year, therefore the marginal distri-bution for primary and general election voters fluctuates year-to-year. This could introduce bias into my estimates. Some may argue that if turnout in the primary election vacillates, it is not a good measure of a district’s primary constituency. I would argue otherwise. Hill (2015) and Sides et al. (Forthcoming) demonstrate that the demographic characteristics and ideological predispositions of voters participating in primary elections do not vary widely across years. Per Fenno (1978), primary voters should be the most dedicated individuals within a constituency. Therefore, while we may see variability in turnout, we should not see too much variability in the types of voters participating.
Third, defining primary and general election constituencies as only those people who voted in an election requires making an assumption about who candidates pay attention to when campaigning. However, incumbent attention allocation and responsiveness has been widely tested in the literature finding that representatives are more responsive to partisan voters than non-voters (Fenno 1978; Clinton 2006; Bafumi and Herron 2010; Bartels 2016).
Disaggregation by Primary and General Electorate
Including an indicator variable for a subunit of interest in the preceding response model allows for estimates to be disaggregated after the postratification step of MrsP by that sub-unit. For instance, including an indicator variable for state and congressional district allows for estimates to be disaggregated at the congressional district level. To disaggregate by primary electorate for each party at the congressional district level requires an indicator variable for a respondent’s state and congressional district, as well as a dummy variable for the partisan primary he voted in.
As stated previously, I use party primary participation as a proxy for whether an in-dividual is part of a party’s primary electorate. Catalist LLC’s Validated Voter Database provides more thorough information about voter participation than traditional survey data through the aggregation of state voter files. The dataset provides complete information for whether a respondent is a voter or non-voter.
voter voted in; data availability varies with a state’s type of primary institution. States with closed and semi-closed primary institutions require voters to register with a party to par-ticipate in the primary election. For these states I place voters in the primary constituency corresponding to their party registration. For independents in semi-closed systems and all voters in open systems, I do not know for certain in which party’s primary a voter partici-pated. Additionally, several states with semi-closed systems do not disclose party registra-tion informaregistra-tion in their voter files. In these instances, I assume a voter participates in the party primary matching their party registration. In the absence of party registration, I use the Catalist partisanship propensity score as a substitute. Much like the ideological extrem-ity score, the partisanship propensextrem-ity score uses covariates in the Catalist, LLC database to predict an individual’s partisan affiliation. I assume a voter participates in the party primary most closely matching their party propensity score.
To estimate the ideological extremity of general electorates, I use general election vote choice as a proxy for whether an individual is part of a party’s general election constituency. Unlike primary electorate affiliation, I cannot make strong inferences for a voter’s general electorate affiliation because it is impossible to know for certain which candidate a person voted for. I classify voters as Democrat or Republican following the same procedure ex-plained above, using an individual’s party registration whenever available or the Catalist party propensity score to infer vote choice.
Results
Fig. 1:Republican Party Primary Electorate Ideology by Congressional District
Figure 1 depicts ideological estimates for Republican primary electorates, Figure 2 de-picts ideological estimates for Democratic primary electorates. The scale for ideological extremity in Figure 1 ranges from 0-50, with 0 being the most conservative. The scale for ideological extremity in Figure 2 ranges from 50-100, with 100 being the most liberal. In Figure 1, the districts with the most conservative primary electorates for the Republican party are shaded lighter. In Figure 2, the districts with the most liberal primary electorates for the Democratic party are shaded darker. The most notable findings in these figures are (1) the ideological variation between states, and (2) the ideological variation within states. Also worth noting is the variation in ideological extremity between Republican and Demo-cratic district estimates. Republican districts seem to exhibit a higher level of variation in extremity, while Democratic districts seem to be more consistently moderate. If ideological extremity in primary constituencies varies by district as well as within and across states, evaluating primary voters on aggregate will not capture ideological heterogeneity. These estimates demonstrate the necessity for a direct measure of primary constituency ideology in evaluations of primary voter influence on candidate behavior.
Comparing primary electorate estimates across districts with similar electoral character-istics demonstrates how indirect measures for district ideology gloss over this ideological heterogeneity. Finding similar results to Hill (2015), there is no consistency in the ideolog-ical extremity of districts in states with the same type of primary institution. Comparing Nebraska and Georgia, both states with open systems, the Republican primary electorates in these states clearly have different ideological distributions. States like Florida — which has a closed system — and Georgia – which has an open system — have comparable levels of ideological extremity in their primary constituencies despite having different types of primary election laws.
Fig. 3:Democratic Party Primary Electorate vs. General Electorate Ideological Estimates
constituency ideology is plotted on the y-axis, general electorate ideology is plotted on the x-axis. Similar to Figures 1 and 2, ideological extremity is on a 0-100 scale with 0 being the most conservative and 100 being the most liberal.
Figure 3 plots Democratic primary and general electorates; districts falling above the reference line have partisan primary electorates that are more ideologically extreme than the general electorate. Figure 4 plots Republican primary and general electorates; districts falling below the reference line have primary electorates that are more ideologically ex-treme than the general electorate. In most cases for Democratic constituencies, the primary electorate is more ideologically extreme than the general electorate. This relationship is less prevalent when examining Republican constituencies. These mixed findings could be motivated by several different factors.
First, in this analysis I am looking at a small sample of cases, I may observe greater vari-ation when I estimate the remaining districts for the 2012 congressional election. Second, to estimate relative primary and general electorate extremity I require marginal distributions for bothparty’s primary electorates and both party’s general electorates. In other words, if any party’s primary race is uncontested in a district I cannot create an estimate for the general electorate’s ideological extremity. This means that those districts for which I can produce estimates may be innately more competitive; both parties ran contested primary elections in these districts. In turn, I may expect constituencies in this subset of districts to be generally more moderate. Third, estimating primary constituency ideology in the year of a presidential election could moderate my estimates. Creating estimates for the 2014 pri-mary and general election would remove the influence of presidential pripri-mary participation.
Discussion & Next Steps
incumbent behavior and candidate emergence to vary across contexts.
Continuing to work on this project, my next step is to conduct robustness checks for my estimates. Hill (2015) notes, “Without clear benchmarks, it is hard to evaluate the pro-cedure outside of the statistical theory that demonstrates that both hierarchical models and post-stratification improve the validity of...estimates to corresponding population statistics.” (Appendix, 6). Using CCES survey data I will estimate an IRT model to predict individual-level ideology as a function of survey policy questions as an attempt to validate my measure. This IRT model will allow for the evaluation of Catalist LLC’s synthetic ideology score in capturing actual voter predispositions. Even if certain districts do not have enough data to produce estimates, I can compare my Catalist-based estimates to those districts that could be estimated with CCES data.
REFERENCES
Abramowitz, Alan. 2008. “Don’t Blame Primary Voters for Polarization.”The Forum5(4).
Aranson, Peter H. and Peter C. Ordeshook. 1972. Spatial strategies for sequential elections. InProbability Models of Collective Decision Making, ed. RG Niemi and HF Weisberg. Columbus, Ohio: Charles E. Merrill Publishing Company.
Bafumi, Joseph and Michael C. Herron. 2010. “Leapfrog Representation and Extremism: A Study of American Voters and Their Members in Congress.”American Political Science Review104(3):519–542.
Bartels, Larry M. 2016. Unequal democracy: The political economy of the new gilded age. Princeton University Press.
Boatright, Robert G. 2014. Getting Primaried: The Changing Politics of Congressional Primary Chal- lenges. University of Michigan Press.
Brady, David W., Hahrie Han and Jeremy C. Pope. 2007. “Primary Elections and Candi-date Ideology: Out of Step with the Primary Electorate?” Legislative Studies Quarterly
32(1):79–105.
Butler, Daniel M. 2009. “The Effect of the Size of Voting Blocs on Incumbents’ Roll-Call Voting and the Asymmetric Polarization of Congress.”Legislative Studies Quarterly
34(3):297–318.
Carson, Jamie L., Jason M. Roberts and Rachel Surminsky. 2017. “The Nationalization of Congressional The Nationalization of Congressional Elections.” Presented at the Annual Meeting of the Midwest Political Science Assocatiation.
Carson, Jamie, Michael H. Crespin, Carrie P. Eaves and Emily O. Wanless. 2012. “Con-stituency Congruency and Candidate Competition in Primary Elections for the U.S. House.”State Politics and Policy Quarterly12(2).
Clinton, Joshua. 2006. “Representation in Congress: constituents and roll calls in the 106th House.”The Journal of Politics68(2).
Fleisher, Richard and John R. Bond. 2004. “The shrinking middle in the US Congress.”
British Journal of Political Science34(3):429–451.
Gelman, Andrew and Thomas C. Little. 1997. “Poststratification Postratification into Many Categories Using Hierarchical Logistic Regression.”Survey Research23(2):127–35.
Grofman, Bernard. 2004. “Downs and Two-Party Covergence.”Annual Review of Political Science7:25–46.
Hall, Andrew. 2015. “What Happens When Extremists Win Primaries?”American Political Science Review109(1):18–42.
Hall, Andrew and James M. Snyder. 2015. “Candidate Ideology and Electoral Success.”.
Hersh, Eitan D. 2015. Hacking the Electorate: How Campaigns Perceive Voters. Cam-bridge University Press.
Hetherington, Marc J., Bruce Larson and Suzanne Globetti. 2003. “The Redistricting Cycle and Strategic Candidate Decisions in U.S. House Races.”Journal of Politics65(4):1221– 1234.
Hill, Seth J. 2015. “Institution of nomination and the policy ideology of primary elec-torates.”Quarterly Journal of Political Science10(4).
Hill, Seth J. and Chris Tausanovitch. 2017. “Southern realignment, party sorting, and the polarization of American primary electorates, 1958-2012.”Public Choice(1-26).
Hirano, Shigeo, James M. Snyder Jr., Stephen Ansolabehere and John Mark Hansen. 2010. “Primary Elections and Partisan Polarization in the U.S. Congress.”Quarterly Journal of Political Science5.
Jacobson, Gary C. 1989. “Strategic Politicians and the Dynamics of U.S. House Elections, 1946-1986.”American Political Science Review83(3):773–93.
Jacobson, Gary C. 2009.The Politics of Congressional Elections. 7th edition ed. New York: Longman.
Kousser, Thad, Jeffrey B. Lewis and Seth E. Masket. 2007. “Ideological adaptation? The survival instinct of threatened legislators.”Journal of Politics69(3):828–843.
Lawless, Jennifer L. and Kathryn Pearson. 2008. “The Primary Reason for Women’s Under-representation? Reevaluating the Conventional Wisdom.”Journal of Politics70(1):67– 82.
Lax, Jeffrey R. and Justin H. Phillips. 2009a. “Gay Rights in the States: Public Opinion and Policy Responsiveness.”American Journal of Political Science103(3):367–86.
Lax, Jeffrey R. and Justin H. Phillips. 2009b. “How Should We Estimate Public Opinion in the States?”American Journal of Political Science53(1):107–21.
Leemann, Lucas and Fabio Wasserfallen. 2017. “Extending the Use and Prediction Preci-sion of Subnational Public Opinion Estimation.”American Journal of Political Science
61(4):1003–1022.
Maestas, Cherie D., Sarah A. Fulton, L. Sandy Maisel and Walter J. Stone. 2006. “When to risk It? Institutions, ambitions, and the decision to run for the U.S. House.”American Political Science Review100(2):195–208.
McGhee, Eric, Seth Masket, Boris Shor, Steven Rogers and Nolan McCarty. 2014. “A Primary Cause of Partisanship? Nomination Systems and Legislator Ideology.”American Journal of Political Science58(2).
Miller, Warren E. and Donald E. Stokes. 1963. “Constituency Influence in Congress.”
American Political Science Review1(57):45–56.
Owen, Guillermo and Bernard Grofman. 2006. “Two-stage electoral competition in two-party contests: persistent divergence of two-party positions.” Social Choice and Welfare
26(3):547–569.
Park, David K., Andrew Gelman and Joseph Bafumi. 2004. “Bayesian Multilevel Estima-tion with PoststratificaEstima-tion: State-Level Estimates from NaEstima-tional Polls.”Political Analy-sis12(4):375–85.
Porter, Rachel and Sarah A. Treul. 2018. “The Primary Decision: Candidate Emergence and Success in Congressional Primaries.” Presented at the Annual Meeting of the Midwest Political Science Assocatiation.
voting.”American Political Science Review83(1):93–121.
Roemer, J.E. 2001. Political Competition: Theory and Applications. Harvard University Press.
Sides, John, Chris Tausanovitch, Lynn Vavreck and Christopher Warshaw. Forthcoming. “On the Representativeness of Primary Electorates.”British Journal of Political Science
.
Sulkin, Tracy. 2005. Issue Politics in Congress. Cambridge University Press.
Tausanovitch, Chris and Christopher Warshaw. 2013. “Measuring Constituent Policy Pref-erences in Congress, State Legislatures, and Cities.”Journal of Politics75(2):330–342.
Thomsen, Danielle M. 2014. “Ideological Moderates Won’t Run: How Party Fit Matters for Partisan Polarization in Congress.”Journal of Politics76(3).
Thomsen, Danielle M and Andrew Hall. 2018. “Who Punishes Extremist Nominees? Can-didate Ideology and Turning Out the Base in US Elections.”American Political Science Review.
Vavreck, Lynn. 2007. “The exaggerated eects of advertising on turnout: The dangers of self-reports.”Quarterly Journal of Political Science2(4):325–343.
